Does chatbot empathy actually make users ask more questions?

The paper separates two things a chatbot's empathy might change: how much users type, and whether they do the work of health inquiry. "Communicative acts" are borrowed from patient-clinician research: asking questions, expressing concerns, and making assertive responses. In a 2 × 2 × 3 within-subjects experiment (N=48) crossing the modality of empathetic expression (Verbal, Visual, Multimodal) with conversational context (General, Sensitive, Mental Health), verbal and multimodal empathy "significantly increased reply length," but communicative acts "were largely shaped by conversational context." The Sensitive context triggered more question-asking. The Mental Health context led to heightened concerns, assertive responses, and unprompted information disclosure.

The authors read this as two different processes. Longer replies under verbal and multimodal empathy may be "conversational politeness or social reciprocity" rather than "proactive information-seeking or assertive behaviors," so empathic cues keep users engaged and typing without making them better informants. Communicative acts, by contrast, follow "the complexity and urgency of their health conditions." The paper calls active participation "fundamentally a problem-driven coping mechanism rather than a reaction to the AI's social behaviors." On this account users ask, worry, and volunteer context because the situation demands it, not because the chatbot was warm.

This bears on the assumption that a more empathetic interface yields richer disclosure. The vault's note on Do empathetic questions serve two completely separate functions? separates what a question does linguistically from its emotional effect. This paper adds a behavioral counterpart: whether a user asks a question at all is set by context more than by the empathic display. It also complements Why can't conversational AI agents take the initiative?, which places missing initiative in the model. Here the variable initiative sits with the user, and the model's social cues turn out to be a weak lever on it. It also sits beside Why do robots outperform chatbots in therapy despite identical language models?, where the medium mattered. The outcomes differ (distress there, communicative acts here), so the two do not directly conflict, but neither supports the idea that expressive form alone moves user behavior.

The excerpt leaves a good deal open. It gives no effect sizes, no account of how acts were coded, and no test of the politeness reading. That reading is offered as what the finding "suggested," and any evidence for it presumably lies in the qualitative findings the excerpt does not show. It lists two three-level factors but labels the design 2 × 2 × 3 without saying what the remaining factor is. It also does not say whether the context conditions differed in more than topic. A within-subjects sample of 48 in a study setting cannot show that the pattern holds with real health worries. What follows at this strength is modest: if the goal is to draw out questions and disclosure, the paper points toward context-sensitive design, which is its own stated implication, and away from tuning empathic tone alone.